We propose to add independent pseudo quantization noise to model parameters during training to approximate the effect of a quantization operator. This method, DiffQ, is differentiable both with respect to the unquantized parameters, and the number of bits used. Given a single hyper-parameter expressing the desired balance between the quantized model size and accuracy, DiffQ can optimize the number of bits used per individual weight or groups of weights, in a single training. We experimentally verify that our method outperforms state-of-the-art quantization techniques on several benchmarks and architectures for image classification, language modeling, and audio source separation. For instance, on the Wikitext-103 language modeling benchmark, DiffQ compresses a 16 layers transformer model by a factor of 8, equivalent to 4 bits precision, while losing only 0.5 points of perplexity. Code is available at: https://github.com/facebookresearch/diffq

This paper proposes a strategy to assess the robustness of different machine learning models that involve natural language processing (NLP). The overall approach relies upon a Search and Semantically Replace strategy that consists of two steps: (1) Search, which identifies important parts in the text; (2) Semantically Replace, which finds replacements for the important parts, and constrains the replaced tokens with semantically similar words. We introduce different types of Search and Semantically Replace methods designed specifically for particular types of machine learning models. We also investigate the effectiveness of this strategy and provide a general framework to assess a variety of machine learning models. Finally, an empirical comparison is provided of robustness performance among three different model types, each with a different text representation.

Inefficient traffic control may cause numerous problems such as traffic congestion and energy waste. This paper proposes a novel multi-agent reinforcement learning method, named KS-DDPG (Knowledge Sharing Deep Deterministic Policy Gradient) to achieve optimal control by enhancing the cooperation between traffic signals. By introducing the knowledge-sharing enabled communication protocol, each agent can access to the collective representation of the traffic environment collected by all agents. The proposed method is evaluated through two experiments respectively using synthetic and real-world datasets. The comparison with state-of-the-art reinforcement learning-based and conventional transportation methods demonstrate the proposed KS-DDPG has significant efficiency in controlling large-scale transportation networks and coping with fluctuations in traffic flow. In addition, the introduced communication mechanism has also been proven to speed up the convergence of the model without significantly increasing the computational burden.

Despite all the challenges and limitations, vision-based vehicle speed detection is gaining research interest due to its great potential benefits such as cost reduction, and enhanced additional functions. As stated in a recent survey [1], the use of learning-based approaches to address this problem is still in its infancy. One of the main difficulties is the need for a large amount of data, which must contain the input sequences and, more importantly, the output values corresponding to the actual speed of the vehicles. Data collection in this context requires a complex and costly setup to capture the images from the camera synchronized with a high precision speed sensor to generate the ground truth speed values. In this paper we explore, for the first time, the use of synthetic images generated from a driving simulator (e.g., CARLA) to address vehicle speed detection using a learning-based approach. We simulate a virtual camera placed over a stretch of road, and generate thousands of images with variability corresponding to multiple speeds, different vehicle types and colors, and lighting and weather conditions. Two different approaches to map the sequence of images to an output speed (regression) are studied, including CNN-GRU and 3D-CNN. We present preliminary results that support the high potential of this approach to address vehicle speed detection.

Position encoding in transformer architecture provides supervision for dependency modeling between elements at different positions in the sequence. We investigate various methods to encode positional information in transformer-based language models and propose a novel implementation named Rotary Position Embedding(RoPE). The proposed RoPE encodes absolute positional information with rotation matrix and naturally incorporates explicit relative position dependency in self-attention formulation. Notably, RoPE comes with valuable properties such as flexibility of being expand to any sequence lengths, decaying inter-token dependency with increasing relative distances, and capability of equipping the linear self-attention with relative position encoding. As a result, the enhanced transformer with rotary position embedding, or RoFormer, achieves superior performance in tasks with long texts. We release the theoretical analysis along with some preliminary experiment results on Chinese data. The undergoing experiment for English benchmark will soon be updated.

In general, an experimental environment for deep learning assumes that the training and the test dataset are sampled from the same distribution. However, in real-world situations, a difference in the distribution between two datasets, domain shift, may occur, which becomes a major factor impeding the generalization performance of the model. The research field to solve this problem is called domain generalization, and it alleviates the domain shift problem by extracting domain-invariant features explicitly or implicitly. In recent studies, contrastive learning-based domain generalization approaches have been proposed and achieved high performance. These approaches require sampling of the negative data pair. However, the performance of contrastive learning fundamentally depends on quality and quantity of negative data pairs. To address this issue, we propose a new regularization method for domain generalization based on contrastive learning, self-supervised contrastive regularization (SelfReg). The proposed approach use only positive data pairs, thus it resolves various problems caused by negative pair sampling. Moreover, we propose a class-specific domain perturbation layer (CDPL), which makes it possible to effectively apply mixup augmentation even when only positive data pairs are used. The experimental results show that the techniques incorporated by SelfReg contributed to the performance in a compatible manner. In the recent benchmark, DomainBed, the proposed method shows comparable performance to the conventional state-of-the-art alternatives. Codes are available at https://github.com/dnap512/SelfReg.

This paper presents a linear dynamical operator described in terms of a rational transfer function, endowed with a well-defined and efficient back-propagation behavior for automatic derivatives computation. The operator enables end-to-end training of structured networks containing linear transfer functions and other differentiable units {by} exploiting standard deep learning software. Two relevant applications of the operator in system identification are presented. The first one consists in the integration of {prediction error methods} in deep learning. The dynamical operator is included as {the} last layer of a neural network in order to obtain the optimal one-step-ahead prediction error. The second one considers identification of general block-oriented models from quantized data. These block-oriented models are constructed by combining linear dynamical operators with static nonlinearities described as standard feed-forward neural networks. A custom loss function corresponding to the log-likelihood of quantized output observations is defined. For gradient-based optimization, the derivatives of the log-likelihood are computed by applying the back-propagation algorithm through the whole network. Two system identification benchmarks are used to show the effectiveness of the proposed methodologies.

Computation and Data Reuse is critical for the resource-limited Convolutional Neural Network (CNN) accelerators. This paper presents Universal Computation Reuse to exploit weight sparsity, repetition, and similarity simultaneously in a convolutional layer. Moreover, CoDR decreases the cost of weight memory access by proposing a customized Run-Length Encoding scheme and the number of memory accesses to the intermediate results by introducing an input and output stationary dataflow. Compared to two recent compressed CNN accelerators with the same area of 2.85 mm^2, CoDR decreases SRAM access by 5.08x and 7.99x, and consumes 3.76x and 6.84x less energy.

We propose a novel loss for generative models, dubbed as GRaWD (Generative Random Walk Deviation), to improve learning representations of unexplored visual spaces. Quality learning representation of unseen classes (or styles) is crucial to facilitate novel image generation and better generative understanding of unseen visual classes (a.k.a. Zero-Shot Learning, ZSL). By generating representations of unseen classes from their semantic descriptions, such as attributes or text, Generative ZSL aims at identifying unseen categories discriminatively from seen ones. We define GRaWD by constructing a dynamic graph, including the seen class/style centers and generated samples in the current mini-batch. Our loss starts a random walk probability from each center through visual generations produced from hallucinated unseen classes. As a deviation signal, we encourage the random walk to eventually land after t steps in a feature representation that is hard to classify to any of the seen classes. We show that our loss can improve unseen class representation quality on four text-based ZSL benchmarks on CUB and NABirds datasets and three attribute-based ZSL benchmarks on AWA2, SUN, and aPY datasets. We also study our loss's ability to produce meaningful novel visual art generations on WikiArt dataset. Our experiments and human studies show that our loss can improve StyleGAN1 and StyleGAN2 generation quality, creating novel art that is significantly more preferred. Code will be made available.

Emotion dynamics modeling is a significant task in emotion recognition in conversation. It aims to predict conversational emotions when building empathetic dialogue systems. Existing studies mainly develop models based on Recurrent Neural Networks (RNNs). They cannot benefit from the power of the recently-developed pre-training strategies for better token representation learning in conversations. More seriously, it is hard to distinguish the dependency of interlocutors and the emotional influence among interlocutors by simply assembling the features on top of RNNs. In this paper, we develop a series of BERT-based models to specifically capture the inter-interlocutor and intra-interlocutor dependencies of the conversational emotion dynamics. Concretely, we first substitute BERT for RNNs to enrich the token representations. Then, a Flat-structured BERT (F-BERT) is applied to link up utterances in a conversation directly, and a Hierarchically-structured BERT (H-BERT) is employed to distinguish the interlocutors when linking up utterances. More importantly, a Spatial-Temporal-structured BERT, namely ST-BERT, is proposed to further determine the emotional influence among interlocutors. Finally, we conduct extensive experiments on two popular emotion recognition in conversation benchmark datasets and demonstrate that our proposed models can attain around 5\% and 10\% improvement over the state-of-the-art baselines, respectively.